arXiv subjects

Yumiao Zhao

Publications and source records attributed to Yumiao Zhao.

9 recordsLinked to original sources

arXiv2026

FFSlim: An Efficient and Lightweight Format for Multi-modal Data Storage and Retrieval

With the rapid expansion of large-scale media-text corpora, multi-modal datasets increasingly require efficient storage and retrieval. Existing formats such as Files, TDP, and FFRecord work adequately for uni-modal data but expose fundamental limitations in multi-modal settings, including storage redundancy, massive small-file overheads, cache-unfriendly layouts, and heavy index structures. These issues jointly inflate storage and memory usage and make I/O the dominant bottleneck in real training workloads. We present FFSlim, a lightweight format for storing and retrieving multi-modal data. FFSlim improves storage efficiency and loading throughput through three components: a unified file format that removes media duplication and avoids small-file proliferation; an adaptive retrieval mechanism that enables low-overhead pair-level access and accelerates repeated media loading; and a redundancy detection and aggregation module that converts existing datasets into the FFSlim layout. The experimental results demonstrate that FFSlim achieves 2.07x and 8.26x higher data loading and write throughput on average than the strongest baseline, with minimal storage and index overhead. Consequently, these underlying I/O accelerations enable FFSlim to reduce end-to-end training time by 5.36%-14.18% across seven diverse multi-modal models.

cs.PF↗

arXiv2026

Janus: Disaggregating Attention and Experts for Scalable MoE Inference

Serving large Mixture-of-Experts (MoE) models is challenging because of their large memory footprints, heterogeneous resource demands, and highly dynamic inference workloads. Most existing MoE inference systems deploy the entire model as a monolithic unit, forcing attention and MoE layers to share the same resource configuration despite their different scaling behaviors and resource bottlenecks. Such coarse-grained provisioning leads to resource inefficiency and suboptimal performance. We present JANUS, a scalable and resource-efficient MoE inference system built around three key principles. First, JANUS disaggregates attention and MoE layers onto separate GPU worker pools, enabling independent resource provisioning for the two layer types, and uses an adaptive two-phase communication mechanism for low-latency data exchange. Second, because MoE-layer execution is often memory-bound and highly sensitive to activated-expert imbalance, JANUS introduces a lightweight, microsecond-scale activation scheduler that balances per-layer activated experts across MoE instances to reduce inference latency. Third, JANUS employs a fine-grained, SLO-aware resource scaling scheme that jointly selects attention resources, MoE resources, and expert placement to minimize GPU cost under token-level SLOs. Evaluation shows that JANUS improves per-GPU throughput by up to 4.7x over state-of-the-art MoE inference baselines while satisfying token-level latency SLOs.

cs.DC↗

arXiv2026

MQAdapter: Multi-Modal Quantum Adapter for Coarse-to-Fine VLM Fine-tuning

Large-scale Vision-Language Models have demonstrated impressive transfer learning capabilities across a wide range of tasks. For few-shot classification, we observe that VLMs exhibit a notable ability to filter candidate categories and thus achieve high Top-K accuracy. However, they often struggle with fine-grained discrimination among visually similar categories, resulting in unsatisfactory Top-1 performance, as shown in Figure 1. Existing studies on VLM adapters generally focus on global alignment between visual and textual representations in the feature space, but fail to exploit semantically similar categories to refine fine-grained visual representations. Based on these observations, we propose a novel coarse-to-fine VLM fine-tuning approach for few-shot learning that leverages quantum computation, termed the Multi-Modal Quantum Adapter (MQAdapter). Specifically, MQAdapter first retrieves the Top-K category candidates most similar to the input image and uses them as semantic anchors. It then employs a cross-modal quantum learning mechanism to refine visual features under the guidance of these anchors. The core of this mechanism is the encoding of visual and textual features into quantum states. By leveraging quantum entanglement and superposition in a high-dimensional Hilbert space, MQAdapter effectively models higher-order cross-modal interactions, producing more discriminative representations than traditional Euclidean adapters. MQAdapter is parameter-efficient and can be integrated with various existing fine-tuning algorithms to achieve further performance gains. Evaluations on 15 datasets demonstrate the effectiveness of MQAdapter while requiring fewer trainable parameters.

cs.CV↗

arXiv2026

Beyond Low-Rank: Low-Rank Sparse Prompting via Spiking Neural Network and Prompt Factorization

Visual Prompting (VP) has emerged as an efficient paradigm for adapting large-scale pre-trained vision models to downstream tasks by incorporating learnable prompts at the input level. However, existing VP methods typically employ dense pixel-level prompts, which often suffer from redundant perturbations, limited generalization and energy inefficiency. To overcome these limitations, we propose to integrate brain-inspired spiking learning into visual prompt learning tasks. As we know that spiking neuron can perform inexpensive information processing by transmitting the input data into discrete spike trains and return sparse outputs. Inspired by this, we propose \textbf{Lo}w-\textbf{R}ank visual \textbf{S}pike \textbf{P}rompting (LoRSP), a novel framework that learns dynamic low-rank sparse visual prompts naturally via a Spiking neuron learning mechanism. The core idea of LoRSP is to exploit the brain-inspired sparse firing mechanism of spiking neurons to generate pixel-level sparse prompt for each instance. To be specific, we first construct a series of prompt factors via low-rank factorization to capture distinct prompt subspaces. These prompt factors are then fed into an SNN architecture, which performs the integrate-and-fire process to emit spikes. As a result, our LoRSP generates a \emph{sparse} visual prompt while maintaining the low-rank constraint. This design enables instance-specific selective prompting, leading to more compact and robust adaptation across diverse downstream tasks. Extensive experiments on five heterogeneous vision backbones and multiple benchmarks demonstrate that LoRSP achieves competitive performance while requiring fewer tunable parameters compared to existing VP methods.

cs.CV↗

arXiv2025

Waltz: Temperature-Aware Cooperative Compression for High-Performance Compression-Based CSDs

Data compression is widely adopted for modern solid-state drives (SSDs) to mitigate both storage capacity and SSD lifetime issues. Researchers have proposed compression schemes at different system layers, including device-side solutions like CCSDs ( c ompression-based c omputational SSDs) and compression supported by host-side, like F2FS (flash-friendly file system). We conduct quantitative studies to understand how host-side and device-side compression schemes affect the temperature and performance of SSD-based storage systems. From our experiments, device-side compression, facilitated by a hardware compression engine, can raise the temperature of CCSDs to intolerable levels, resulting in throttling and service shutdown. In contrast, host-side compression causes software-stack overhead, which often results in large performance degradation and resource consumption. To ensure efficient data compression with high performance and better temperature control, we propose Waltz, a temperature-aware cooperative compression method that schedules (de)compression tasks at the host and device sides by monitoring device temperature. Furthermore, we introduce two variants (Waltzs and Waltzp) for space and performance optimization, respectively. Waltz is implemented within F2FS, achieving high performance while extending SSD lifetime and preventing overheating-induced in-flight shutdowns.

cs.PF↗

arXiv2025

ConZone+: Practical Zoned Flash Storage Emulation for Consumer Devices

To facilitate the understanding and efficient enhancement of software and hardware design for consumer-grade zoned flash storage, ConZone is proposed as the first emulator designed to model the resource constraints and architectural features typical of such systems. It incorporates essential components commonly deployed in consumer-grade devices, including limited logical to physical mapping caches, constrained write buffers, and hybrid flash media management. However, ConZone cannot be mounted with the file system due to the lack of in-place update capability, which is required by the metadata area of F2FS. To improve the usability of the emulator, ConZone+ extends ConZone with support for a block interface. We also provide a script to help the deployment and introduces several enhancements over the original version. Users can explore the internal architecture of consumer-grade zoned flash storage and integrate their optimizations with system software using ConZone+. We validate the accuracy of ConZone+ by comparing a hardware architecture representative of consumer-grade zoned flash storage and comparing it with the state-of-the-art. In addition, we conduct several case studies using ConZone+ to investigate the design of zoned storage and explore the inadequacies of the current file system.

cs.AR↗

arXiv2025

RARO: Reliability-aware Conversion with Enhanced Read Performance for QLC SSDs

Quad-level cell (QLC) flash offers significant benefits in cost and capacity, but its limited reliability leads to frequent read retries, which severely degrade read performance. A common strategy in high-density flash storage is to program selected blocks in a low-density mode (SLC), sacrificing some capacity to achieve higher I/O performance. This hybrid storage architecture has been widely adopted in consumer-grade storage systems. However, existing hybrid storage schemes typically focus on write performance and rely solely on data temperature for migration decisions. This often results in excessive mode switching, causing substantial capacity overhead. In this paper, we present RARO (Reliability-Aware Read performance Optimization), a hybrid flash management scheme designed to improve read performance with minimal capacity cost. The key insight behind RARO is that much of the read slowdown in QLC flash is caused by read retries. RARO triggers data migration only when hot data resides in QLC blocks experiencing a high number of read retries, significantly reducing unnecessary conversions and capacity loss. Moreover, RARO supports fine-grained multi-mode conversions (SLC-TLC-QLC) to further minimize capacity overhead. By leveraging real-time read retry statistics and flash characteristics, RARO mitigates over-conversion and optimizes I/O performance. Experiments on the FEMU platform demonstrate that RARO significantly improves read performance across diverse workloads, with negligible impact on usable capacity.

cs.AR↗

arXiv2025

Fine-Grained VLM Fine-tuning via Latent Hierarchical Adapter Learning

Adapter-based approaches have garnered attention for fine-tuning pre-trained Vision-Language Models (VLMs) on few-shot classification tasks. These methods strive to develop a lightweight module that better aligns visual and (category) textual representations, thereby enhancing performance on downstream few-shot learning tasks. However, existing adapters generally learn/align (category) textual-visual modalities via explicit spatial proximity in the underlying embedding space, which i) fails to capture the inherent one-to-many associations between categories and image samples and ii) struggles to establish accurate associations between the unknown categories and images. To address these issues, inspired by recent works on hyperbolic learning, we develop a novel Latent Hierarchical Adapter (LatHAdapter) for fine-tuning VLMs on downstream few-shot classification tasks. The core of LatHAdapter is to exploit the latent semantic hierarchy of downstream training data and employ it to provide richer, fine-grained guidance for the adapter learning process. Specifically, LatHAdapter first introduces some learnable `attribute' prompts as the bridge to align categories and images. Then, it projects the categories, attribute prompts, and images within each batch in a hyperbolic space, and employs hierarchical regularization to learn the latent semantic hierarchy of them, thereby fully modeling the inherent one-to-many associations among categories, learnable attributes, and image samples. Extensive experiments on four challenging few-shot tasks show that the proposed LatHAdapter consistently outperforms many other fine-tuning approaches, particularly in adapting known classes and generalizing to unknown classes.

cs.CV↗

arXiv2024

HeGraphAdapter: Tuning Multi-Modal Vision-Language Models with Heterogeneous Graph Adapter

Adapter-based tuning methods have shown significant potential in transferring knowledge from pre-trained Vision-Language Models to the downstream tasks. However, after reviewing existing adapters, we find they generally fail to fully explore the interactions between different modalities in constructing task-specific knowledge. Also, existing works usually only focus on similarity matching between positive text prompts, making it challenging to distinguish the classes with high similar visual contents. To address these issues, in this paper, we propose a novel Heterogeneous Graph Adapter to achieve tuning VLMs for the downstream tasks. To be specific, we first construct a unified heterogeneous graph mode, which contains i) visual nodes, positive text nodes and negative text nodes, and ii) several types of edge connections to comprehensively model the intra-modality, inter-modality and inter-class structure knowledge together. Next, we employ a specific Heterogeneous Graph Neural Network to excavate multi-modality structure knowledge for adapting both visual and textual features for the downstream tasks. Finally, after HeGraphAdapter, we construct both text-based and visual-based classifiers simultaneously to comprehensively enhance the performance of the CLIP model. Experimental results on 11 benchmark datasets demonstrate the effectiveness and benefits of the proposed HeGraphAdapter.

cs.CV↗